A free, local, open-source alternative to Wispr Flow.
Hold a key, talk, and polished text lands in whatever text box you're in.
Fast, private, open source. Windows, macOS and Linux.

See it in action:

https://github.com/user-attachments/assets/78b5c7a9-8fa1-4c4d-b439-02c90f8bacef

---

Ochre is a free alternative to subscription dictation apps. Speech recognition runs on your machine by default. If you prefer the cloud, bring your own API key and pay cents an hour instead of a monthly fee. An optional AI pass cleans up what you said: it removes "um"s, applies your "no wait, make that Tuesday"s, and fixes punctuation, without ever rewriting your words or answering your questions.

Highlights

• Fast. Text appears ~150 ms after you release the key (median, measured on a Ryzen 7 9800X3D, CPU only), even with the machine busy. Phrases are transcribed while you talk; the mic is always warm, so the first syllable is never clipped.

• Accurate locally. NVIDIA Parakeet TDT 0.6B (v3 "Ultra", int8) runs on ONNX Runtime and supports 25 European languages, with punctuation and casing built in. The choice is benchmarked against Whisper, Qwen3-ASR, Granite, Moonshine and others.

• Your keyboard, your rules. Configurable Voice key. Hold to talk, double-tap to lock recording on, then tap to stop, and Esc to cancel. Text is typed with native keystrokes, or pasted with your clipboard restored for long text.

• Hands-free. Say "transcribe", dictate, then say "transcribe stop" to insert, "transcribe send" to insert and press Enter, or "transcribe cancel" to discard. The custom wake-word model is trained in this repo and false-wakes about 0.4 times per hour (the model itself is licensed for non-commercial use, see License).

• Optional cleanup. Local or cloud (OpenAI, Google, Groq, OpenRouter, Anthropic). Locally it runs our own fine-tunes, Ochre Refine 4B / 2B / 0.8B (Qwen3.5, trained on real recognizer output). It removes fillers, applies self-corrections, fixes punctuation and writes numbers the way you would, without rewording you or answering your questions. On 773 dictations graded blind by three LLM judges, it passes 80% (4B), 71% (2B) and 64% (0.8B) (how it was trained and measured). The default, auto, picks by hardware: 4B on a GPU with 6 GB or more (or Apple Silicon with 16 GB), 2B on a smaller GPU, 0.8B on CPU only. The models are on Hugging Face under Apache-2.0. A guard falls back to your raw words if a model ever misbehaves.

• Bring your own key. Pick a connector, paste one key, done:

  Connector	Speech-to-text	Cleanup	Rough cost
  Local (default)	Parakeet on your CPU/GPU	Ochre Refine / Quill (optional)	free
  OpenAI	gpt-transcribe / live streaming	GPT-5.6 Luna (minimal reasoning)	cents per hour
  Google	Gemini 3.5 Transcribe	Gemini 3.5 Flash-Lite	cents per hour
  Soniox, Deepgram, Groq, ElevenLabs, AssemblyAI	mix and match under Advanced

• Private by design. No telemetry, audio never touches disk, and history is local text you can turn off. API keys live in your OS keychain.

Install

Download the installer for your system from the latest release:

macOS and Linux support is newer than Windows: if something doesn't work on your setup, please open an issue.

Or build from source; it's one command. You can also clone the repo and ask your coding agent (Claude Code, Codex, Cursor…) to "install Ochre": AGENTS.md tells it how.

Windows (PowerShell):

git clone https://github.com/jordan-gibbs/ochre; cd ochre
./scripts/install.ps1

macOS / Linux:

git clone https://github.com/jordan-gibbs/ochre && cd ochre
./scripts/install.sh

The script installs Rust if needed, plus the OS prerequisites (WebView2 on Windows; webkit2gtk, libayatana-appindicator and ALSA dev packages on Linux). It then builds and installs ochre (on macOS, Ochre.app in /Applications, started for you). On first launch it downloads the speech model (~670 MB) with a progress bar; turn on Start at login in Settings if you want it always running.

macOS. The first-run screen walks you through Microphone, Accessibility and Input Monitoring (an Allow… button each; no restart needed). The microphone goes through the Mac's own voice processing; in a noisy room, choose Voice Isolation in the microphone mode menu (last setup step, or Settings → General). Using fn (Globe) as the Voice key? Set "Press 🌐 key to" to Do Nothing. Details and test results: docs/macos.md.

Linux. On X11 everything works as installed (the Voice key through XInput, typing with xdotool). On Wayland:

• Voice key: Wayland gives apps no global keys, so Ochre reads /dev/input. Join the input group (sudo usermod -aG input $USER, then log out and back in), or bind a keyboard shortcut to ochre toggle in your desktop settings (press once to start, again to stop; no permissions needed).
• Typing: GNOME and KDE have no virtual-keyboard protocol, so Ochre types with ydotool through /dev/uinput. The .deb / .rpm and install.sh add a udev rule that lets the logged-in user use it, and Ochre starts ydotoold itself. Sway, Hyprland and other wlroots desktops use wtype. Text that ydotool can't type (non-ASCII, or a non-US layout) is pasted instead, and your clipboard is put back.
• Tray: Ubuntu shows the tray icon out of the box. On stock GNOME, install the AppIndicator and KStatusNotifierItem Support extension; without it there is no tray icon, and launching Ochre again from the app menu opens Settings.

Use

Everything you dictate is in History (tray → History), so nothing is lost if focus moved.

How it works

key / wake word ─▶ warm mic (pre-roll) ─▶ phrase segmenter ─▶ speech-to-text (local or cloud)
                                              decodes while you talk          │
     injection ◀─ history ◀─ cleanup (optional, guarded) ◀─ hesitation strip ◀┘

One native Rust process (Tauri v2 for the pill, tray and settings). The design rules come from measuring why an earlier app felt slow (docs/latency.md): models stay resident and warm, decoding runs at raised priority without spin-waits, there is no polling, text is typed before anything is saved, and every dictation logs its per-stage timings.

See SPEC.md for the full design and docs/go-checklist.md for the measured release gates.

Development

cargo test                          # all crates
cargo run -p ochre-app              # the app (add -- --demo to drive every UI state)
cargo run -p ochre-stt --example bench --release -- clip.wav   # speech latency
cargo run -p ochre-refine --example bench --release            # cleanup latency

Model training lives in training/: the wake word (training/wakeword) and the cleanup model (training/refine, with a synthetic-speech data pipeline and an eval harness in tools/eval).

Contributing

Bug reports, fixes and measured improvements are welcome. Read CONTRIBUTING.md first: it covers building and testing, the PR flow (fork, branch, PR to main; small focused PRs, one maintainer reviews everything) and what is in scope. Please follow the Code of Conduct, and report security problems privately as described in SECURITY.md.

License

Ochre's source code is released under the MIT License, Copyright (c) 2026 Jordan Gibbs.

Model weights, fonts and bundled binaries keep their own licences, listed in THIRD_PARTY_NOTICES.md: Parakeet is CC-BY-4.0, Qwen3.5, Quill and the Ochre Refine fine-tunes are Apache-2.0, the openWakeWord front end is Apache-2.0, and the fonts are under the SIL Open Font License 1.1.

The one exception is the "transcribe" wake-word model (assets/wake/transcribe.onnx), which is licensed CC BY-NC-SA 4.0: free for non-commercial use. Some of the data it was trained on only allows non-commercial use (openWakeWord licenses its own models the same way). The training code is MIT, and everything else in Ochre works without the wake word.